← All news

Weekly recap

AI News — Week of June 1–7, 2026

#weekly

Microsoft Build 2026 dominated the week — Windows reframed as an agent platform running Microsoft's own models — while Anthropic scaled Project Glasswing and filed to go public, and AI's billing economics moved to center stage.

The week in brief

Microsoft Build 2026 (June 2–3) set the week’s agenda: Windows recast as an agent platform, and Microsoft’s own MAI models moving in to replace OpenAI’s inside GitHub Copilot. Around it, two parallel stories defined the week — vendors building their own frontier models to cut dependence, and the editor turning into a console for supervising swarms of agents.

Biggest stories

  • [highest impact] Microsoft Build makes Windows an agent platform and reveals its homegrown models. The Windows Agent Framework was open-sourced under MIT alongside an Agent Store, WSL 3, and Azure Agent Mesh; Azure AI Foundry Agent Service and the Microsoft Foundry portal hit GA. Microsoft also named Project Polaris / a family of seven MAI models (MAI-Thinking-1, MAI-Code-1) set to replace GPT-4 Turbo as the Copilot default from August — its clearest move yet to reduce OpenAI dependence in its flagship dev tool. (briefing, briefing, official)
  • Anthropic scales Project Glasswing to ~150 critical-infrastructure orgs across 15+ countries, with the first ~50 partners already surfacing 10,000+ high/critical flaws (Cloudflare ~2,000, Mozilla 271 in Firefox 150), and ships Claude Security — a controlled expansion, not GA. (briefing, official)
  • Anthropic confidentially files a draft S-1, getting out ahead of OpenAI, on a ~$965B valuation and reported ~$47B revenue run-rate — a vendor-durability signal for teams standardizing on Claude. (briefing, official)
  • Cognition retires Windsurf, relaunching it as Devin Desktop — an Agent Command Center (a Kanban board of local and cloud agents), a Rust-rewritten Devin Local, and support for the open Agent Client Protocol (ACP) driving Codex, Claude Agent, and OpenCode. Cascade is deprecated July 1. (briefing, official)
  • GitHub Copilot individual plans move to usage-based AI Credits, with base + flex allotments and a new Max tier — the week’s clearest sign that flat-rate AI pricing is giving way to metered economics. (briefing, official)

By area

  • Model releases — Microsoft’s seven-model MAI family and Project Polaris led the week; OpenAI shipped a GPT-Rosalind life-sciences update (31% fewer tokens than GPT-5.5) and confirmed GPT-4.5’s June 27 / o3’s August 26 retirements; DeepSeek made its 75% V4-Pro price cut the permanent floor.
  • Coding agents — Copilot Workspace exited beta with Agent Mode, a standalone Copilot desktop app and a CLI refresh (/rubber-duck, prompt scheduling, voice) all landed at Build; Cursor shipped 3.7 with visual Design Mode and split Teams pricing pools.
  • MCP — the VIPER-MCP study scanned ~39,884 MCP-server repos and surfaced 106 zero-days (67 CVEs assigned), the largest empirical look yet at MCP servers as an attack surface.
  • AI-assisted SDLC — Microsoft IQ unified enterprise grounding across Copilot/Foundry; OpenAI’s GPT-5.5/5.4 and Codex reached GA on Amazon Bedrock; Anthropic formalized its Partner Network; the Linux Foundation announced a Tokenomics Foundation to standardize token-cost telemetry.

Themes

  • Own the model, not just the product. Microsoft’s MAI/Polaris push to displace OpenAI inside Copilot was the sharpest example, but the pattern ran through the week — DeepSeek locking in an aggressive price floor, OpenAI shipping vertical gated models (GPT-Rosalind) — as vendors move from renting frontier capability to controlling it.
  • The IDE becomes a console for supervising agent swarms. Copilot Workspace’s meta-agent, GitHub’s standalone Copilot app orchestrating parallel sessions, and Devin Desktop’s Command Center all reframe the editor from a place you write code into a board where you triage and unblock several long-running agents.
  • AI’s economics turned into first-class infrastructure. GitHub’s AI Credits, Cursor’s split usage pools, the looming Anthropic billing split, DeepSeek’s price floor, and the Linux Foundation’s Tokenomics Foundation all surfaced the same shift: token spend is becoming something teams must budget, meter, and govern like any other cost.

Still watching

  • Gemini 3.5 Pro — promised for June at I/O with no committed date; fresh reporting pegs a 2M-token window, a “Deep Think” mode, and ~$15/$60 per 1M-token pricing. An aggressive price would make it a plausible default to move a coding agent onto. (unconfirmed)
  • Anthropic developer billing split (June 15) — Claude Agent SDK, claude -p, and Claude Code GitHub Actions reportedly move onto a separate credit pool (~$20–$200) metered at full API rates with no rollover; sharpest impact on bursty CI automation. Anthropic still hasn’t detailed it officially. (unconfirmed)
  • MCP spec 2026-07-28 final — stateless core, Tasks, and MCP Apps go stable and clients must validate iss per RFC 9207; in light of the VIPER-MCP CVE wave, the authorization hardening is worth getting ahead of.
  • Project Polaris → Copilot migration (August) & MAI benchmarks — Microsoft named the models but published no independent numbers; the “tuned for GitHub” framing holds up or doesn’t once third parties test MAI-Code-1 on real repositories.
  • xAI Grok V9-Medium (mid-June) — a reported ~1.5-trillion-parameter model aimed at the coding lead, no model card or independent benchmarks yet. (unconfirmed)